The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill interfaces with external data by fetching user-submitted content through the get_form_responses and get_form tools. This creates an attack surface where an external actor could submit a form response containing adversarial instructions intended to hijack the agent's context once the data is read.
Ingestion points: The get_form_responses and get_form tools described in SKILL.md retrieve external, untrusted content.
Boundary markers: The skill does not provide instructions for using delimiters or boundary markers when the agent processes form data.
Capability inventory: The agent has the ability to create, update, and delete forms, as well as perform shell commands via the bun runtime configuration described in the setup section.
Sanitization: There are no specific instructions regarding the sanitization or validation of input received from the Google Forms API.
[COMMAND_EXECUTION]: The setup documentation instructs the user to execute shell commands (bun install, bun run build, bun run token) and configures the agent to run a local script via bun in the .mcp.json file. While standard for MCP (Model Context Protocol) servers, this involves the execution of local code and requires manual oversight during the initial configuration.

google-forms